A noise-robust and cross-domain few-shot fault diagnosis method of rolling bearings based on TFC-FPN

Abstract Accurate fault identification of rolling bearings is crucial for the safe operation of rotating machinery. However, noise interference and frequent changes in operating conditions limit the effectiveness of traditional diagnostic methods in practical applications. To address the issues, we propose a fault diagnosis method combining a time-frequency convolutional feature pyramid network (TFC-FPN) with a pretrained-finetuning transfer strategy, aiming to improve diagnostic accuracy and generalization under high-noise, fluctuating-noise, cross-domain and small-sample scenarios. This approach leverages both time and frequency domain features, along with self-attention pyramid pooling and multi-scale attention fusion, significantly enhancing multiscale fault feature extraction capabilities. For cross-domain diagnostics, the pretrained-finetuning transfer strategy is utilized to mitigate overfitting in scenarios with limited samples across different domains. This approach enhances the model’s generalization ability across varying devices and operational conditions. Experimental results show that the TFC-FPN model outperforms others, demonstrating exceptional diagnostic performance in complex and dynamic environments. Compared to other SOTA models, TFC-FPN shows superior generalization in complex environments, highlighting its broad potential for industrial applications.